Georeferencing a Scanned Map

When you scan a paper map, the file remembers only one thing: which pixel is which colour. The top-left pixel is row 0, column 0; there is no metre, no latitude, nowhere. Georeferencing builds a mathematical bridge from that pixel grid to a real coordinate reference system (CRS) — the framework of numbers, introduced in Module 0, that pins a location to the Earth.

You do it by naming pairs. You point at a place on the scan whose real coordinate you know — a surveyed road junction, a temple, a plot corner — and you tell the software: this pixel is that coordinate. Each such pair is a ground control point (GCP). Give the software enough well-chosen GCPs and it can compute a transform that repositions every remaining pixel, so the whole sheet snaps onto the world.

The quality of a georeference is decided almost entirely by the GCPs, so choose them like a surveyor, not a clicker.

Pick sharp, unambiguous features — road intersections, building corners, bridge ends — not fuzzy things like a tree canopy or a river bend that shifts every monsoon. Spread them to the edges and corners. Four points bunched in the middle leave the sheet free to swing at its margins; points pushed into all four corners lock it down. Use more than the minimum. An affine transform needs only three points, but adding a few extra gives you something priceless: a check. And know where the truth comes from — you can type coordinates from a trusted source, click matching points on an already-georeferenced reference layer (imagery, OSM), or enter values printed in a survey sheet's own grid.

Once the GCPs exist, you choose a transformation — the rule for moving every other pixel. They form a ladder from rigid to rubbery.

A first-order (affine) transform allows only shift, rotation, uniform scale and shear — straight lines stay straight. It is the right default for a flat, well-printed sheet and for most master plans. A second-order polynomial permits gentle bending, useful for a sheet with mild, smooth distortion. A thin-plate spline (rubber-sheeting) bends locally to pass through every GCP exactly — powerful for a crumpled or unevenly stretched old sheet, but dangerous, because it can warp wildly between your points. The rule: choose the simplest transform that meets your accuracy tolerance. More flexibility fits the control points better and can quietly ruin everything in between.

Because you supplied more GCPs than the transform strictly needs, the software can measure how far each point lands from where you said it should be. That miss, per point, is the residual. The root-mean-square (RMS) error rolls all the residuals into a single distance — one honest number for the whole fit, in the units of your target CRS (metres, if you chose a projected system).

Read it critically. A low RMS is necessary but not sufficient: a rubber-sheet transform can drive RMS to near zero and still be garbage between points. Watch the per-point residuals too — one GCP with a huge residual is usually a mis-click or a typo, and deleting or re-placing it often fixes the whole sheet. Decide a tolerance before you start ("good enough that a 2 m feature lands within 2 m") and accept the georeference only when the RMS is under it.

Two Indian wrinkles matter. First, datum. Older Survey of India topographic sheets were drawn on the Everest ellipsoid with a Kalianpur datum, not WGS 84. If you type printed grid coordinates from such a sheet, set the georeferencer's target CRS to the matching legacy system (for example a Kalianpur 1975 India zone) or reproject afterwards — otherwise your sheet will sit tens to hundreds of metres off modern imagery. Second, licence and provenance. Survey of India Open Series Maps are free to Indian residents but non-commercial, and a scan's own accuracy is limited by its source scale — a 1:50,000 sheet cannot honestly resolve a plot boundary. Georeferencing does not add accuracy the paper never had.